Adaptive Measurements
Start at 128 shots. Escalate cumulatively to 256, 512, and 1024 only when the current score is inside the stage’s validation-calibrated margin around the decision threshold.
NeurIPS 2026 · SaTQuML Workshop
Adaptive-ShotHybrid Quantum
Anomaly Detection
Validation-calibrated measurement allocation for Tactile Internet security. AS-VQC begins at 128 shots and adds measurements only when an anomaly score remains close to a fixed decision threshold.
Research companion: explore the published method and results. This static Space does not run a detector or execute quantum circuits.
Framework Overview
Adaptive-Shot Variational Quantum Circuit (AS-VQC) inference treats quantum measurements as a per-record resource. A trained hybrid scorer is reused without retraining or threshold recalibration.
Decision-Level Reliability
Measure both ranking quality and disagreement with the analytic decision at the same threshold. Stable ranking does not necessarily imply stable individual decisions.
Controlled Evaluation
Compare fixed-shot budgets, three adaptive calibration levels, and a budget-shuffled control across random, entity-group-disjoint, and temporal holdouts.
Method and evaluation: paper, Section 3. The “95” in AS-VQC-95 is a validation-error percentile, not a formal 95% test-correctness guarantee.
4,875
CESNET Candidate Records
12
Features / Registered Qubits
15
Trained QNN Checkpoints
1,200
Finite-Shot Policy Realizations
Five checkpoints per holdout. Measurement repetitions are not additional independently trained models. Candidate counts precede training-derived filtering.
Adaptive Evidence, Cached Enforcement
Mirrored telemetry is analyzed off-path. The current Tactile Internet request continues through deterministic cached enforcement while later evidence can inform policy updates.
Off-Path Evidence Plane
Mirrored Aggregate-Flow Telemetry
Training-Derived Feature Preparation
Classical Embedder → Quantum Encoding
12-Qubit VQC · Joint Pauli-Z Readout
128 Cumulative Shots → Classical Head
Outside the Calibrated Threshold Margin?Yes: return evidence · No: acquire more shots
Escalate: 256 → 512 → 1024Reuse accumulated measurements; stop at the cap
Return Anomaly Evidence + Shots Used
→Subsequent
Policy Updates
Policy Updates
On-Path Enforcement Plane
Client / Tactile Request
Policy Enforcement Point
Cached Deterministic Policy Lookup
Grant · Restrict · Step-Up · Deny
Protected Tactile Internet Service
Stopping rule
|p̂S − τ*| > δS,βStop when the finite-shot score’s distance from the fixed threshold exceeds the validation-error margin. Otherwise add only enough shots to reach the next cumulative stage. All remaining records stop at 1024.
Scope: this is the scorer’s architectural role, not a measured full PDP/PEP deployment. Reduced shot counts do not themselves establish haptic-loop latency, physical-device energy savings, or production security.
Architecture and stopping policy: paper, Sections 3.3–3.6.
Measurement Savings & Reliability
Primary policy: AS-VQC-95. Measurement realizations are averaged within each checkpoint, then summarized across five checkpoints per holdout. These are reported paper results, not measurements generated by this Space.
| Holdout | ROC-AUCMean ± SD | Average ShotsMean ± SD per record | Saving vs. Fixed-1024Mean % ± SD in pp | Decision DisagreementMean vs. analytic reference |
|---|---|---|---|---|
| Random | 0.9956 ± 0.0014 | 129.2 ± 0.4 | 87.39% ± 0.04 pp | 0.771% |
| Entity-Group | 0.9968 ± 0.0017 | 276.9 ± 327.3 | 72.96% ± 31.97 pp | 0.409% |
| Temporal | 0.9980 ± 0.0010 | 131.2 ± 2.0 | 87.19% ± 0.20 pp | 0.635% |
Source: paper, Table 1 and Section 4. “pp” denotes percentage points. Saving = 100 × (1 − average shots / 1024); independently rounded values may not reproduce exactly. Disagreement is with analytic inference, not verified attack ground truth.
Retain the tradeoff: Fixed-1024 is more decision-stable than AS-VQC-95. The Group variability is real in the reported matrix: seed 46 averages 862.4 shots per record, while the other four Group checkpoints remain near the 128-shot floor. It is not excluded as an outlier.
Compared with Fixed-128
AS-VQC-95 has lower reported decision disagreement in each holdout. Its primary advantage is at the thresholded-decision level rather than an identified ranking gain.
Matched Measurement Budget
Budget-Shuffled-AS95 preserves the realized shot-budget distribution but randomly reassigns budgets to records. AS-VQC-95 has lower disagreement in all three holdouts, testing targeted allocation rather than quantity alone.
More Conservative Calibration
AS-VQC-99 reports lower mean disagreement than Fixed-512 while averaging 162.6, 432.9, and 207.7 shots for Random, Group, and Temporal. AS-VQC-95 remains the primary policy.
View the complete AS-VQC-95 predictive metrics
| Holdout | ROC-AUC | Average Precision | FPR | ECE · 15 bins |
|---|---|---|---|---|
| Random | 0.9956 ± 0.0014 | 0.9760 ± 0.0050 | 0.0299 ± 0.0042 | 0.0257 ± 0.0062 |
| Entity-Group | 0.9968 ± 0.0017 | 0.9835 ± 0.0092 | 0.0346 ± 0.0312 | 0.0174 ± 0.0038 |
| Temporal | 0.9980 ± 0.0010 | 0.9895 ± 0.0060 | 0.0252 ± 0.0104 | 0.0211 ± 0.0076 |
Mean ± sample SD across five checkpoints. FPR and expected calibration error (ECE) are fractions, not percentages. Labels are training-quantile statistical pseudo-labels.
Experimental Configuration
Hybrid Scorer
12 → 64 → 12 classical embedder; 12-qubit, two-layer VQC with 72 registered parameters and 22 CNOTs; Pauli-Z readout on qubits 0 and 1; 2 → 32 → 2 classical head. The configured readout has three-wire causal support.
Training & Calibration
20 epochs · batch size 32 · Adam at 2 × 10⁻³ · seeds 42–46. Preprocessing uses training data; validation selects checkpoints, thresholds, and adaptive margins. Test labels do not control stopping.
Inference & Controls
Analytic training followed by ideal joint-multinomial sampling. Four fixed budgets, AS-VQC-90/95/99, and a matched-budget shuffled policy reuse the same checkpoints. No physical hardware noise, queueing, or device timing is evaluated.
Benchmark boundary: the 4,875-record CESNET-TimeSeries24 candidate pool contains aggregate network telemetry, not dedicated haptic traffic. Quantile-derived anomaly labels do not establish verified intrusion detection. Validation-error percentiles are empirical calibrations, not formal test-time correctness guarantees.
Configuration and limitations: paper, Section 3 and Section 6. The linked repository contains the experiment workflow; this Space contains no checkpoints or private/raw telemetry.
Citation
Please cite the associated paper when using AS-VQC, its methodology, or the reported measurement–reliability results.
BibTeX · arXiv:2610.05835
@misc{sudipto2026adaptiveshothybridquantum,
title = {Adaptive-Shot Hybrid Quantum Anomaly Detection for Tactile Internet Security: Reliability-Aware Measurement Allocation Under Resource Constraints},
author = {Sudipto, Mubassir Serneabat and Ahmed, Shakil and Khokhar, Ashfaq and Iqbal, Samir M.},
year = {2026},
eprint = {2610.05835},
archivePrefix = {arXiv},
primaryClass = {cs.CR},
doi = {10.48550/arXiv.2610.05835},
url = {https://arxiv.org/abs/2610.05835},
note = {NeurIPS 2026 SaTQuML Workshop; short oral presentation reported in the arXiv comments}
}Workshop presentation status is reported in the arXiv comments; it is not a NeurIPS main-conference acceptance claim.